1. Sweden blocks ChatGPT and Claude for government staff
Sweden blocks ChatGPT and Claude on the computers of its Government Offices starting Monday, October 26, 2026. The Government Offices, the administrative body that supports the prime minister and the cabinet, will steer employees toward Microsoft Copilot, which is the only artificial intelligence tool to have cleared an internal security review. The stated reason for the restriction is the protection of information.
The decision affects the staff who prepare policy, draft documents and handle correspondence for Sweden's executive branch. It does not ban artificial intelligence outright. It picks one product, from one American technology company, and shuts out two others, which makes it a more selective measure than a blanket prohibition would have been.
2. What did the Government Offices say about the risks?
The explanation came from Mikael Wettercrantz, the director of digitalisation at the Government Offices, who set out the reasoning in a message on the staff intranet. He told employees that using AI services without a formal contract can have consequences for the organization's information security.
Wettercrantz also pointed to several examples of what he described as problematic use of open AI services. The concerns he raised were of two kinds. One was the sensitivity of the material that employees type into these tools. The other was the way staff handle the answers that the tools produce, which can be inaccurate or can carry material that should not be reused without checking.
Those two points together describe the practical risk of consumer-grade assistants in an office that deals with confidential government business. A prompt can contain the text of a draft, the details of an internal discussion or the name of a person. Once it is sent, it sits on infrastructure that the government does not control.
3. Why a contract matters for government use of AI
Free or consumer versions of AI assistants send what a user types to servers operated by outside companies. In the case of providers based in the United States, American law allows authorities to compel such companies to hand over data when they are asked to do so. For a foreign government, that is a material consideration, because information entered by a civil servant could in principle become reachable by a third country's agencies.
A formal security agreement changes that picture. It defines what a provider may and may not do with the inputs of government employees, covering matters such as retention, use for model training and access by the provider's own staff. That is the central difference between a vetted enterprise arrangement and an open consumer tool that anyone can reach with a web browser and an email address.
This is the logic behind the distinction the Government Offices have drawn. The question they asked was not which model writes the best answers. It was whether a written commitment exists that governs the handling of what staff enter.
4. How did Microsoft Copilot pass the review?
Copilot passed because the Government Offices examined it and then raised its security settings. In other words, the approval rests on a review combined with configuration choices made by the government itself, rather than on the product in its default form. ChatGPT and Claude did not receive that treatment, and they are the two tools being blocked.
The blocked tools are not wholly out of reach for everyone. Some employees can still use both through an AI platform that the Government Offices built internally. The press office has not said how many people have this access or what restrictions apply to it, so the size of the exception is unknown.
That internal platform illustrates a common approach among organizations that want to give staff access to several models while keeping the traffic under their own control. Access runs through a system the organization manages, rather than through the public websites of the model makers.
5. What separates an enterprise AI tool from a consumer chatbot?
Consumer chatbots are designed for individuals. A person signs up, accepts standard terms of service and begins typing. The organization that employs that person has no say in the terms, cannot audit how the data is handled and often does not know the tool is being used at all. This is the pattern that security teams describe as shadow IT, and AI assistants have made it much more common because they are free, fast and useful.
Enterprise offerings are built around the opposite assumption. The customer organization signs a contract, administrators control which features and models are switched on, and the agreement spells out how inputs are stored and whether they can be used to improve the provider's systems. Administrators can also apply the access controls, logging and data boundaries that the organization already uses for its email and document systems.
For a government office, those controls are not a convenience but a precondition. Public bodies handle personal data, internal deliberations and material that may be classified or otherwise protected by law. A tool that sits outside the organization's security perimeter, and outside its contracts, is difficult to square with those duties.
The remarks about how staff handle AI-generated answers deserve attention alongside the data concerns. Language models can produce fluent text that contains errors, and a rushed employee may paste such text into an official document without verification. In a government setting, a mistake in a briefing note or a draft decision can carry consequences that go well beyond an embarrassing correction.
Restricting the choice of tools does not remove that risk by itself, and Copilot can also produce incorrect output. What a managed deployment can do is make usage visible and bring it under policy, so that guidance, training and review can be attached to the one assistant that staff are allowed to use. That is easier to organize than guidance covering a changing set of public services that nobody has formally assessed.
6. Which other Swedish agencies have taken similar steps?
The Government Offices are not the first Swedish public body to act. The Tax Agency and the National Board of Institutional Care have already imposed comparable bans on consumer AI tools. The Enforcement Authority, which handles debt collection and court-ordered enforcement, is preparing its own decision on the subject.
An internal audit at the Transport Administration pointed to the scale of the problem. It found roughly 10 AI services in use at the agency in a manner it called widespread and risky. The finding shows how quickly employees can adopt such services on their own when no approved alternative is on offer.
Taken together, the measures suggest a pattern across Swedish public administration. Agencies are moving from informal tolerance of AI tools toward a position in which only reviewed products are allowed on work computers.
7. How does this compare with the European Parliament and Italy?
Governments across Europe have been uneasy about the data practices of AI services for several years. Italy banned ChatGPT in 2023, and the worry about where prompts go and who can read them has stayed on the agenda since then.
In February, the European Parliament disabled AI features on all work devices, a step that covered ChatGPT, Claude and Copilot alike. The institution concluded that it could not guarantee the security of the data those tools process. That decision treated the three products in the same way.
Sweden's approach draws a narrower line than the one drawn in Brussels. It approves a single American product, Copilot, and shuts out two others. The difference reflects the Swedish emphasis on whether a product has been reviewed and configured, instead of a general judgment about the category.
8. Is Claude already inside Copilot?
The line between the banned and approved tools is less clear than the announcement suggests, because of a development earlier this year. Microsoft added Claude models to Copilot in early 2026. In the European Union, where Sweden is a member, Anthropic's models are switched off by default, and administrators have to turn them on manually.
The Government Offices have not confirmed whether Claude is enabled in their own Copilot deployment. If it is, staff would be blocked from using Claude directly while still being able to reach its models through the approved assistant. If it is not, the restriction would be cleaner in practice.
The point shows that the choice is as much about the contract and the controls around a service as about the model behind it. The same underlying technology can fall on either side of the line depending on how it is delivered and who is responsible for the data.
9. What the decision signals for AI providers
For the companies behind ChatGPT and Claude, the Swedish decision is a reminder that public-sector customers weigh contracts and controls heavily, and that a strong model is not enough to win access. A provider that wants government staff to use its product has to offer terms that a security review can accept, and that satisfy a buyer worried about foreign legal reach into its data.
For governments, the episode shows the other side of the same trade-off. Choosing one approved assistant simplifies oversight, but it concentrates reliance on a single vendor and leaves open questions about which underlying models that vendor makes available. The unresolved status of Claude inside the approved assistant is one example of how those questions can appear in practice.
10. What comes next for AI rules in the Swedish public sector?
The next visible step is the decision of the Enforcement Authority, which has yet to announce its position. If it follows the Tax Agency and the National Board of Institutional Care, three of the agencies that have drawn attention in this area would have chosen the same direction.
For now, the practical effect is limited to the Government Offices and begins on October 26. Staff there will find ChatGPT and Claude unavailable on their work computers and will be pointed to Copilot. The details that remain open are the number of employees with access through the internal platform and the status of Claude within the approved assistant, and neither has been disclosed.